1/*
  2 * Lockless get_user_pages_fast for MIPS
  3 *
  4 * Copyright (C) 2008 Nick Piggin
  5 * Copyright (C) 2008 Novell Inc.
  6 * Copyright (C) 2011 Ralf Baechle
  7 */
  8#include <linux/sched.h>
  9#include <linux/mm.h>
 10#include <linux/vmstat.h>
 11#include <linux/highmem.h>
 12#include <linux/swap.h>
 13#include <linux/hugetlb.h>
 14
 15#include <asm/cpu-features.h>
 16#include <asm/pgtable.h>
 17
 18static inline pte_t gup_get_pte(pte_t *ptep)
 19{
 20#if defined(CONFIG_64BIT_PHYS_ADDR) && defined(CONFIG_CPU_MIPS32)
 21	pte_t pte;
 22
 23retry:
 24	pte.pte_low = ptep->pte_low;
 25	smp_rmb();
 26	pte.pte_high = ptep->pte_high;
 27	smp_rmb();
 28	if (unlikely(pte.pte_low != ptep->pte_low))
 29		goto retry;
 30
 31	return pte;
 32#else
 33	return ACCESS_ONCE(*ptep);
 34#endif
 35}
 36
 37static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
 38			int write, struct page **pages, int *nr)
 39{
 40	pte_t *ptep = pte_offset_map(&pmd, addr);
 41	do {
 42		pte_t pte = gup_get_pte(ptep);
 43		struct page *page;
 44
 45		if (!pte_present(pte) ||
 46		    pte_special(pte) || (write && !pte_write(pte))) {
 47			pte_unmap(ptep);
 48			return 0;
 49		}
 50		VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 51		page = pte_page(pte);
 52		get_page(page);
 53		SetPageReferenced(page);
 54		pages[*nr] = page;
 55		(*nr)++;
 56
 57	} while (ptep++, addr += PAGE_SIZE, addr != end);
 58
 59	pte_unmap(ptep - 1);
 60	return 1;
 61}
 62
 63static inline void get_head_page_multiple(struct page *page, int nr)
 64{
 65	VM_BUG_ON(page != compound_head(page));
 66	VM_BUG_ON(page_count(page) == 0);
 67	atomic_add(nr, &page->_count);
 68	SetPageReferenced(page);
 69}
 70
 71static int gup_huge_pmd(pmd_t pmd, unsigned long addr, unsigned long end,
 72			int write, struct page **pages, int *nr)
 73{
 74	pte_t pte = *(pte_t *)&pmd;
 75	struct page *head, *page;
 76	int refs;
 77
 78	if (write && !pte_write(pte))
 79		return 0;
 80	/* hugepages are never "special" */
 81	VM_BUG_ON(pte_special(pte));
 82	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 83
 84	refs = 0;
 85	head = pte_page(pte);
 86	page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 87	do {
 88		VM_BUG_ON(compound_head(page) != head);
 89		pages[*nr] = page;
 90		if (PageTail(page))
 91			get_huge_page_tail(page);
 92		(*nr)++;
 93		page++;
 94		refs++;
 95	} while (addr += PAGE_SIZE, addr != end);
 96
 97	get_head_page_multiple(head, refs);
 98	return 1;
 99}
100
101static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
102			int write, struct page **pages, int *nr)
103{
104	unsigned long next;
105	pmd_t *pmdp;
106
107	pmdp = pmd_offset(&pud, addr);
108	do {
109		pmd_t pmd = *pmdp;
110
111		next = pmd_addr_end(addr, end);
112		/*
113		 * The pmd_trans_splitting() check below explains why
114		 * pmdp_splitting_flush has to flush the tlb, to stop
115		 * this gup-fast code from running while we set the
116		 * splitting bit in the pmd. Returning zero will take
117		 * the slow path that will call wait_split_huge_page()
118		 * if the pmd is still in splitting state. gup-fast
119		 * can't because it has irq disabled and
120		 * wait_split_huge_page() would never return as the
121		 * tlb flush IPI wouldn't run.
122		 */
123		if (pmd_none(pmd) || pmd_trans_splitting(pmd))
124			return 0;
125		if (unlikely(pmd_huge(pmd))) {
126			if (!gup_huge_pmd(pmd, addr, next, write, pages,nr))
127				return 0;
128		} else {
129			if (!gup_pte_range(pmd, addr, next, write, pages,nr))
130				return 0;
131		}
132	} while (pmdp++, addr = next, addr != end);
133
134	return 1;
135}
136
137static int gup_huge_pud(pud_t pud, unsigned long addr, unsigned long end,
138			int write, struct page **pages, int *nr)
139{
140	pte_t pte = *(pte_t *)&pud;
141	struct page *head, *page;
142	int refs;
143
144	if (write && !pte_write(pte))
145		return 0;
146	/* hugepages are never "special" */
147	VM_BUG_ON(pte_special(pte));
148	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
149
150	refs = 0;
151	head = pte_page(pte);
152	page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
153	do {
154		VM_BUG_ON(compound_head(page) != head);
155		pages[*nr] = page;
156		if (PageTail(page))
157			get_huge_page_tail(page);
158		(*nr)++;
159		page++;
160		refs++;
161	} while (addr += PAGE_SIZE, addr != end);
162
163	get_head_page_multiple(head, refs);
164	return 1;
165}
166
167static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
168			int write, struct page **pages, int *nr)
169{
170	unsigned long next;
171	pud_t *pudp;
172
173	pudp = pud_offset(&pgd, addr);
174	do {
175		pud_t pud = *pudp;
176
177		next = pud_addr_end(addr, end);
178		if (pud_none(pud))
179			return 0;
180		if (unlikely(pud_huge(pud))) {
181			if (!gup_huge_pud(pud, addr, next, write, pages,nr))
182				return 0;
183		} else {
184			if (!gup_pmd_range(pud, addr, next, write, pages,nr))
185				return 0;
186		}
187	} while (pudp++, addr = next, addr != end);
188
189	return 1;
190}
191
192/*
193 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
194 * back to the regular GUP.
195 */
196int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
197			  struct page **pages)
198{
199	struct mm_struct *mm = current->mm;
200	unsigned long addr, len, end;
201	unsigned long next;
202	unsigned long flags;
203	pgd_t *pgdp;
204	int nr = 0;
205
206	start &= PAGE_MASK;
207	addr = start;
208	len = (unsigned long) nr_pages << PAGE_SHIFT;
209	end = start + len;
210	if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
211					(void __user *)start, len)))
212		return 0;
213
214	/*
215	 * XXX: batch / limit 'nr', to avoid large irq off latency
216	 * needs some instrumenting to determine the common sizes used by
217	 * important workloads (eg. DB2), and whether limiting the batch
218	 * size will decrease performance.
219	 *
220	 * It seems like we're in the clear for the moment. Direct-IO is
221	 * the main guy that batches up lots of get_user_pages, and even
222	 * they are limited to 64-at-a-time which is not so many.
223	 */
224	/*
225	 * This doesn't prevent pagetable teardown, but does prevent
226	 * the pagetables and pages from being freed.
227	 *
228	 * So long as we atomically load page table pointers versus teardown,
229	 * we can follow the address down to the page and take a ref on it.
230	 */
231	local_irq_save(flags);
232	pgdp = pgd_offset(mm, addr);
233	do {
234		pgd_t pgd = *pgdp;
235
236		next = pgd_addr_end(addr, end);
237		if (pgd_none(pgd))
238			break;
239		if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
240			break;
241	} while (pgdp++, addr = next, addr != end);
242	local_irq_restore(flags);
243
244	return nr;
245}
246
247/**
248 * get_user_pages_fast() - pin user pages in memory
249 * @start:	starting user address
250 * @nr_pages:	number of pages from start to pin
251 * @write:	whether pages will be written to
252 * @pages:	array that receives pointers to the pages pinned.
253 *		Should be at least nr_pages long.
254 *
255 * Attempt to pin user pages in memory without taking mm->mmap_sem.
256 * If not successful, it will fall back to taking the lock and
257 * calling get_user_pages().
258 *
259 * Returns number of pages pinned. This may be fewer than the number
260 * requested. If nr_pages is 0 or negative, returns 0. If no pages
261 * were pinned, returns -errno.
262 */
263int get_user_pages_fast(unsigned long start, int nr_pages, int write,
264			struct page **pages)
265{
266	struct mm_struct *mm = current->mm;
267	unsigned long addr, len, end;
268	unsigned long next;
269	pgd_t *pgdp;
270	int ret, nr = 0;
271
272	start &= PAGE_MASK;
273	addr = start;
274	len = (unsigned long) nr_pages << PAGE_SHIFT;
275
276	end = start + len;
277	if (end < start || cpu_has_dc_aliases)
278		goto slow_irqon;
279
280	/* XXX: batch / limit 'nr' */
281	local_irq_disable();
282	pgdp = pgd_offset(mm, addr);
283	do {
284		pgd_t pgd = *pgdp;
285
286		next = pgd_addr_end(addr, end);
287		if (pgd_none(pgd))
288			goto slow;
289		if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
290			goto slow;
291	} while (pgdp++, addr = next, addr != end);
292	local_irq_enable();
293
294	VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
295	return nr;
296slow:
297	local_irq_enable();
298
299slow_irqon:
300	/* Try to get the remaining pages with get_user_pages */
301	start += nr << PAGE_SHIFT;
302	pages += nr;
303
304	down_read(&mm->mmap_sem);
305	ret = get_user_pages(current, mm, start,
306				(end - start) >> PAGE_SHIFT,
307				write, 0, pages, NULL);
308	up_read(&mm->mmap_sem);
309
310	/* Have to be a bit careful with return values */
311	if (nr > 0) {
312		if (ret < 0)
313			ret = nr;
314		else
315			ret += nr;
316	}
317	return ret;
318}